AI spending shifts toward model routing and open-weight systems

Artificial intelligence companies are moving from a contest centered on the largest models toward systems that choose among models by task, cost and operating environment. The shift could pressure the pricing power of OpenAI, Anthropic and other premium model providers as companies scrutinize AI budgets, CNBC reported.

Perplexity Chief Executive Aravind Srinivas told CNBC that the model itself is becoming only one part of the product. He said value is shifting to the software layer that selects a model, connects it to tools and applies it to company data.

That approach is often described as model routing. A system may assign routine customer service work to a lower-cost model, send a difficult coding request to a more capable model, or keep an internal workflow on an open model controlled by the company. The commercial question is no longer limited to which model scores highest on a benchmark. Buyers are also weighing latency, data access, deployment location and usage cost.

“The model alone is no longer the product,” Srinivas told CNBC. “It is the harness, the orchestration system that puts the model inside a very capable harness and pairs the model with a lot of tools.”

Open models challenge premium economics

Perplexity this week showed a new system for its computer-use product built around GLM 5.2, an open model from China’s Z.ai, according to CNBC. The design lets a cheaper model complete more steps and calls on a stronger model only when the task requires it.

Open-weight models can be downloaded, adjusted and operated by companies on their own infrastructure. As those models improve, they offer an alternative to paying premium rates for proprietary models from large AI labs. Tokens, the units of data that AI systems process and produce, are central to that cost structure because providers often charge based on usage.

Benchmark general partner Peter Fenton told CNBC that open-weight systems could account for more than 90% of tokens generated within 18 to 24 months, and possibly by year-end. He said inference margins at frontier model companies could face pressure as customers run “good enough” open-weight models without the markup charged by proprietary providers.

Fenton also said the economics are only part of the case. Smaller models adapted to a narrow task can sometimes respond faster and perform better than larger general-purpose systems, he told CNBC.

Control becomes a buyer priority

Benchmark has invested in Ollama, which provides tools for developers and enterprises to download, run and manage open models. Ollama Chief Executive Jeff Morgan told CNBC that corporate buyers focus heavily on operating control, including where a model runs and how it is managed.

Morgan said Ollama is used by more than 85% of Fortune 500 companies, including businesses in aviation, insurance and health care. He told CNBC that many companies begin with smaller models running close to their own data and later expand to larger open models.

The trend also carries policy implications for the United States. CNBC reported that some of the leading open-weight models are being developed by Chinese labs, including Z.ai and DeepSeek. Srinivas told CNBC that U.S. support for open models would help make AI cheaper and more widely available to small businesses in America and allied countries.

The operating model could affect the data center buildout now underway across the technology sector. Srinivas told CNBC that some AI work may eventually run locally on consumer or business devices, while harder tasks are sent to cloud systems with more powerful models. That would leave a hybrid structure rather than removing demand for data centers.

For investors and operators, the issue is whether the largest AI labs can preserve pricing power as open models improve and corporate customers become more selective about which systems they use for each workload.

This story draws on original reporting from CNBC.